Efficient Federated Learning with Encrypted Data Sharing for Data-Heterogeneous Edge Devices

Guodong Fan; Hangyu Li; HongYue Wu; Shizhan Chen; Zhen Zhang; Zhiyong Feng

REVIEW 4 major objections 6 minor 32 references

Efficient Federated Learning with Encrypted Data Sharing for Data-Heterogeneous Edge Devices

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FedEDS is claimed to accelerate federated learning convergence and improve accuracy under data heterogeneity by having edge clients share stochastic-layer-encrypted data and by annealing local epochs over communication rounds.

desk verdict Solid empirical FL data-sharing technique with a privacy claim that doesn't survive contact with the paper's own equations. read the letter →

arxiv 2506.20644 v2 pith:H7UH7F7Q submitted 2025-06-25 cs.LG

Hangyu Li , Hongyue Wu , Guodong Fan , Zhen Zhang , Shizhan Chen , Zhiyong Feng This is my paper · ORCID

classification cs.LG

keywords federatedlearningedgecomputingdataheterogeneityencryptedsharingknowledgetransferstochasticlayerpiecewiseepochannealingnon-IID

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Federated learning slows down and loses accuracy when the data on different edge devices is unevenly distributed, and the usual fix still requires many slow round-trips to a central server. The paper proposes FedEDS, a plug-in that lets each edge client train a small encryptor whose outputs are shared with other clients so every client learns from the others' data without exposing raw samples. It claims this accelerates convergence, improves top-1 accuracy on CIFAR-10 and FashionMNIST, and can be dropped into FedAvg, FedProx, and FedNova. The mechanism relies on a frozen stochastic layer inside the client model that makes the shared data unreadable to humans.

What carries the argument

The central object is the data encryptor trained with a frozen stochastic layer. Equation (3) trains $\phi_k$ so that $f(g(x; \phi_k); \theta^g_k, \theta^c_k, \theta^l_k)$ matches the label $y$, while the frozen stochastic layer parameters $\theta^l_k$ add noise perturbations during forward propagation; the encryptor must learn to counter those perturbations, and the resulting images are visually unrecognizable yet class-informative. Sharing these encrypted datasets (Eq. 4) plus the stochastic-layer parameters is what transfers knowledge between clients. The second mechanism is the piecewise epoch annealing schedule (Eq. 5), which starts local training with $E_{max}$ epochs and decays to $E_{min}$ at turning rounds $T^\alpha_{turn}$ and $T^\beta_{turn}$, cutting the number of central aggregation rounds.

What would settle it

Run a model-inversion or membership-inference attack against the shared encrypted datasets $(g(x; \phi_k), \theta^l_k)$: if original images can be visually reconstructed or membership can be reliably inferred, the paper's privacy claim fails. Also, compare FedEDS against data-free federated distillation baselines that share no data at all; if those reach the same accuracy in the same number of rounds, the encrypted sharing is not what drives the reported gains.

Watch

Extended reading notes

Core claim

On its own terms, FedEDS claims to solve two problems at once: data heterogeneity and central-server communication latency. Each client trains a U-Net encryptor $g(x; \phi_k)$ so that encrypted data, when passed through the client model with its frozen stochastic layer, preserves the class information of the original image; it then publishes $g(x; \phi_k)$ along with the stochastic-layer parameters $\theta^l_k$. Other clients train on these encrypted datasets with a KL-divergence loss while updating their own feature extractor, plus their local raw data, with weights $\lambda_c$ and $\lambda_{dis}$ decaying over rounds. A piecewise epoch annealing schedule reduces local epochs as training progresses. Experiments report that FedEDS plus FedAvg reaches the same accuracy in far fewer communication rounds and achieves higher final accuracy, with the largest gains under high heterogeneity ($\alpha=0.1$).

Load-bearing premise

The privacy guarantee collapses if the stochastic-layer-based 'encrypted' data can be inverted or linked to the original private samples; the paper provides no adversary model or cryptographic analysis to rule this out.

Editorial extensions

If this is right

FedEDS can be layered onto FedAvg, FedProx, and FedNova as a plug-in, so existing federated learning deployments may gain accuracy without changing their aggregation logic.
Communication rounds to reach a target accuracy drop substantially; for example, on CIFAR-10 with 20 clients and high heterogeneity, FedEDS plus FedAvg reaches 66% accuracy in 60 rounds versus 186 rounds for FedAvg alone.
The accuracy gain is largest when data is most heterogeneous, which is exactly the regime where unmodified federated learning performs worst.
The total client-side training cost is roughly unchanged for small client counts and is reduced for large client counts, because encryptor training happens once during initialization.
The ablation study attributes most of the improvement to encrypted data knowledge transfer rather than to the epoch annealing schedule, with the combination performing best.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's 'encryption' is not cryptographic; nothing in the mechanism prevents a malicious or curious client from running model-inversion or membership-inference attacks on the shared datasets, so the privacy claim should be tested empirically before deployment.
Since each round a client trains on only one uniformly sampled other client's encrypted data, the method is closest to a randomized pairwise distillation; the convergence proof relies on assumptions about the encrypted data distribution that the paper inherits from existing analysis rather than establishing.
If edge-to-edge links are not significantly faster than edge-to-cloud links, the extra encrypted-data broadcast could outweigh the reduction in central rounds; a cost model comparing both link types would settle whether FedEDS helps in a given deployment.
The approach could be combined with secure aggregation or differential privacy to make the privacy guarantee stronger, but that would add noise or computation that the current experiments do not measure.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

I read FedEDS with some care. The mechanism is genuinely new: each client trains a UNet to make images that survive the frozen stochastic layer, shares those images plus soft labels with the other clients, and those clients train on the mix. The ablation shows the data-sharing part, not the epoch schedule, is what drives the gains. And the gains look consistent: across two datasets, three heterogeneity levels, and 5/10/20 clients, FedEDS beats the three baselines and reaches target accuracies in far fewer rounds. The write-up is reproducible, with code, hyperparameters, and a clean experimental protocol. That's more than most papers at this depth.

The soft spot is the privacy story. Eq. (4) defines the shared object as (g(x; phi_k), f(g(x); theta_g, theta_c, theta_l)). The second component is the sender's soft prediction. The encryptor is trained so that that soft prediction matches the true class. So the shared data directly encodes class membership. If the labels themselves are sensitive - and in many FL applications they are - the 'privacy-preserving' description is wrong before you even worry about inverting the UNet. And no one inverts the UNet: there is no threat model, no attack evaluation, no analysis of the deterministic transform's one-wayness. Visual unrecognizability is not a privacy guarantee. This is not a minor omission; it's the central advertised benefit.

The convergence theorem is likewise thinner than it looks. The proof is not in the paper; it's borrowed from [27], and the paper doesn't verify that the assumptions hold for their loss with encrypted data and swapped stochastic layers. State it as a conjecture, not a theorem, until the argument is filled in.

Minor: no error bars or multiple seeds. The gains are a few percent, so this matters, but the pattern across many settings is consistent enough that I don't doubt the core result.

For a reader: this is a useful paper for the FL data-sharing/knowledge-transfer crowd, but it should be read as a method paper with a questionable privacy label, not as a privacy technology. It deserves a serious referee, though a reviewer should push hard on the privacy model and the convergence claim.

My recommendation: send to peer review; the mechanism and empirical evidence are worth the referee's time, but major revision is likely on the privacy framing.

Referee Report

4 major / 6 minor

Summary. The paper proposes FedEDS, a federated learning method for edge devices that combines encrypted data sharing among clients with a piecewise epoch annealing schedule. Each client trains a UNet-based data encryptor using its local model and a frozen stochastic layer, then broadcasts an 'encrypted' version of its local dataset (encrypted inputs plus the sender model's soft labels) to other clients. Recipients train on their own private data plus one randomly sampled encrypted dataset from another client, with the sender's stochastic layer swapped in. The method is designed as a plug-in for FedAvg, FedProx, and FedNova, and the paper reports accuracy and communication-round improvements on CIFAR-10 and FashionMNIST under Dirichlet-partitioned heterogeneity with 5, 10, and 20 clients. A convergence theorem is stated but not proved, with the proof attributed to an external reference. The empirical tables consistently show accuracy gains and reduced rounds for FedEDS across configurations.

Significance. If the empirical results and the privacy assumptions hold, FedEDS would be a practically attractive drop-in enhancement for federated learning in edge settings: it is compatible with several standard algorithms, the code is made publicly available, and the reported communication-round reductions are substantial, especially for K=20. The idea of using a model's stochastic layer to train a data encryptor that produces transferable but visually unrecognizable inputs is interesting and clearly presented. However, the paper's privacy claim is a core part of its framing, and it is not backed by any formal privacy analysis or attack evaluation; the shared soft labels alone are a plausible leakage channel. The convergence theorem is asserted rather than demonstrated. The empirical evidence also lacks variance estimates, which matters because several reported gains are small (e.g., less than 0.3 percentage points in Table I).

major comments (4)

[Section IV-A, Eq. (4)] The privacy claim that FedEDS shares data 'without compromising data privacy' is load-bearing and unsupported. In Eq. (4), the shared object D^F_k consists of g(x; φ_k), a deterministic UNet transform of the private input, paired with the sender model's soft prediction f(g(x; φ_k); θ_g, θ_c, θ_l). Because Eq. (3) trains the encryptor so that this soft prediction is close to the true class, the shared soft labels directly encode class-membership information; if labels are sensitive, privacy is violated even if the images look unrecognizable. The paper provides no adversary model, no formal privacy guarantee (e.g., differential privacy), and no membership-inference or inversion attack evaluation. Visual unrecognizability in Fig. 3 is not a privacy guarantee. The authors should either drop the privacy claim, restrict it to a perceptual-obfuscation claim, or provide a concrete threat model and empirical/formal evidence that the shared dataset does not leak class membership or allow input reconstruction.
[Section V-B, Theorem 1] The convergence theorem is stated as following from 'the mathematical proof in [27]' without including a proof or a verification that the assumptions of [27] hold for the FedEDS objective. The actual FedEDS loss in Eq. (8) involves uniform random sampling of one other client's encrypted dataset, the time-varying weights λ_c and λ_dis of Eqs. (9)-(10), and the swapping of stochastic layers θ_l_i; it is not immediate that the four stated assumptions (smoothness, bounded scalar, unbiased gradient/bounded variance, bounded dissimilarity) are satisfied by this non-convex, data-dependent objective. In particular, the gradient with respect to θ_g^k on the L_dis term depends on the sampled client i*, so the unbiased-gradient assumption requires a careful treatment of the sampling distribution. The bound in Eq. (13) should either be proved in the manuscript or the authors should state precisely which conditions of [27] are inherited and why.
[Section VI-B, Tables I-II] The central empirical claim of consistent accuracy improvement is based on single runs with no error bars, no multiple seeds, and no statistical significance tests. Several reported gains are very small (e.g., CIFAR-10, K=5, α=10: 75.11 vs. 75.34 for FedAvg vs. FedEDS+FedAvg in Table I, a 0.23-point difference), and it is not possible to tell whether such differences are reproducible or within run-to-run noise. Given that the paper's main positive claim is 'consistently demonstrates outstanding performance,' the experiments should report mean and standard deviation over at least 3-5 independent runs, or otherwise justify that the reported numbers are stable.
[Section VI-C and Algorithm 3] The communication-efficiency analysis counts communication rounds to the central server but does not account for the one-time broadcast of each client's full encrypted dataset D^F_k to all other K-1 clients. Since D^F_k is comparable in size to the local dataset, this peer-to-peer transfer adds O(K * |D_k|) data movement per client. The paper asserts that edge-to-edge links are fast and that 'the mutual transmission of encrypted data by FedEDS does not bring significant additional communication delays,' but no measurement or even rough bandwidth/latency estimate is provided. To substantiate the headline claim of reducing the impact of central-server delays, the authors should quantify the encrypted-data-sharing overhead in practical edge scenarios.

minor comments (6)

[Abstract and Section I] Typo: 'a new federated learning scheme on edge devices that called' should be 'that is called' or 'that we call'; similar phrasing issues appear elsewhere.
[Algorithm 4] Algorithm 4 calls the encrypted-data generation routine 'FedFeatGenerate' while Algorithm 1 defines 'FedEncryptedDataGenerate'; the names should be consistent.
[Table II caption] The dataset name is misspelled as 'FashionMINST' in the caption; it should be 'FashionMNIST'.
[Section IV-B, Eq. (5)] The floor notation in Eq. (5) is used but not explained; a brief note that brackets denote the floor function would improve readability.
[Section III, Eq. (1)] The notation p_k = n_k / n is standard, but the definition of L_k(θ_k) uses the same symbol n for the total dataset size and n_k for local size; this is clear from context but could be restated for precision.
[Section IV-C, Eq. (8)] The constraint 'subject to i* ∼ U({1,...,K} \ {k})' is written inside the min expression; it should be placed outside the optimization to avoid suggesting that the client optimizes over the random index.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FedEDS's empirical gains are measured, and its convergence theorem cites an external proof rather than reducing to the method's own definitions.

full rationale

The paper's central empirical claims are measured outcomes of controlled experiments (Tables I-VII), not quantities derived from the method's own definitions or from fitted parameters, so they cannot reduce to inputs by construction. The only formal result, Theorem 1, is explicitly conditional on standard assumptions and is anchored to an external proof in [27]; borrowing that proof is a citation of independent support, not a circular import, and the theorem is not needed to make the empirical claims. Eq. (4) does define the shared 'encrypted' label as the sender model's own soft prediction, so the shared data are teacher outputs; however, the downstream accuracy of a different client model is measured, and the improvement is an empirical finding rather than a consequence of the definition. The privacy guarantee in the abstract and Section IV-A is asserted without an adversary model or attack evaluation, but an unsupported privacy claim is a correctness risk, not a circular derivation. Hence no circular step is identified.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The method depends on several hand-chosen hyperparameters for the epoch schedule and loss weighting, standard FL assumptions for the theoretical bound, and an untested domain assumption that the encrypted data preserves privacy. The only new component, the data encryptor, is a learned model with no independent evidence of its privacy properties.

free parameters (6)

Emax = 5
Maximum number of local epochs in early communication rounds, chosen by hand in Section VI-A.
Emin = 1
Minimum number of local epochs in later rounds, chosen by hand.
T_turn_alpha = 1
First turning point in the piecewise epoch annealing schedule (Eq. 5).
T_turn_beta = 3
Second turning point in the epoch schedule.
m = 3
Controls the decay rate of lambda_dis and lambda_c (Eqs. 9-10).
epsilon = 0.01
Threshold below which lambda_dis is set to 0.

assumptions (4)

standard math Standard FL assumptions (Lipschitz smoothness, bounded variance, bounded dissimilarity) hold for the FedEDS loss.
Assumptions 1-4 in Section V-B are standard in FL theory and are invoked without verification for the modified loss with encrypted data.
standard math The convergence bound in Theorem 1 follows from the proof in reference [27].
The paper does not provide its own proof and simply states the theorem based on [27], so the result is not self-contained.
domain assumption Encrypted data generated by the data encryptor does not reveal private information.
The privacy claim is not formally proven or empirically tested; it is assumed that the stochastic-layer perturbation is sufficient.
domain assumption The stochastic layer parameters of other clients can be safely incorporated into a client's model during training.
The knowledge transfer objective in Eq. (6) assumes that replacing the stochastic layer with another client's version yields a valid learning signal.

invented entities (1)

Data encryptor model (UNet)
purpose: Transforms private local data into encrypted images that are shared with other clients while preserving label knowledge.
The data encryptor is a newly introduced component whose privacy and utility properties are only demonstrated in the paper's experiments; no external validation or formal guarantee is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Efficient Federated Learning with Encrypted Data Sharing for Data-Heterogeneous Edge Devices." pith.science (2026). https://pith.science/paper/H7UH7F7Q

@misc{pith2026250620644,
  author       = {Pith},
  title        = {Pith review of: Efficient Federated Learning with Encrypted Data Sharing for Data-Heterogeneous Edge Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H7UH7F7Q}},
  note         = {Machine review of arXiv:2506.20644}
}

read the original abstract

As privacy protection gains increasing importance, more models are being trained on edge devices and subsequently merged into the central server through Federated Learning (FL). However, current research overlooks the impact of network topology, physical distance, and data heterogeneity on edge devices, leading to issues such as increased latency and degraded model performance. To address these issues, we propose a new federated learning scheme on edge devices that called Federated Learning with Encrypted Data Sharing(FedEDS). FedEDS uses the client model and the model's stochastic layer to train the data encryptor. The data encryptor generates encrypted data and shares it with other clients. The client uses the corresponding client's stochastic layer and encrypted data to train and adjust the local model. FedEDS uses the client's local private data and encrypted shared data from other clients to train the model. This approach accelerates the convergence speed of federated learning training and mitigates the negative impact of data heterogeneity, making it suitable for application services deployed on edge devices requiring rapid convergence. Experiments results show the efficacy of FedEDS in promoting model performance.

Figures

Figures reproduced from arXiv: 2506.20644 by the authors.

**Figure 2.** In Step 1 of the figure, the client model is trained normally using its local dataset, during which the forward propagation of the model skips the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** The first row of the figure showcases original images from the CIFAR-10 [26] dataset, while the second and the third row visualizes the impact [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: The figure illustrates the process of optimizing the model on client [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: The visualizations of the data partitioning utilized in the experiments are presented, where the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 14 canonical work pages

[27]

Fake It Till Make It: Federated Learning with Consensus-Oriented Generation

R. Ye, Y . Du, Z. Ni, S. Chen, and Y . Wang, “Fake it till make it: Federated learning with consensus-oriented generation,” arXiv preprint arXiv:2312.05966, 2023

work page Pith review arXiv 2023
[1]

Communication-efficient learning of deep networks from decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial intelligence and statistics . PMLR, 2017, pp. 1273– 1282

2017
[2]

Federated machine learning: Concept and applications,

Q. Yang, Y . Liu, T. Chen, and Y . Tong, “Federated machine learning: Concept and applications,” ACM Transactions on Intelligent Systems and Technology (TIST), vol. 10, no. 2, pp. 1–19, 2019

2019
[3]

Q. Yang, L. Fan, and H. Yu, Federated learning: Privacy and incentive. Springer Nature, 2020, vol. 12500

work page 2020
[4]

Fedpe: Adaptive model pruning-expanding for federated learning on mobile devices,

L. Yi, X. Shi, N. Wang, J. Zhang, G. Wang, and X. Liu, “Fedpe: Adaptive model pruning-expanding for federated learning on mobile devices,” IEEE Transactions on Mobile Computing , 2024

work page 2024
[5]

Fedmp: Federated learning through adaptive model pruning in heterogeneous edge computing,

Z. Jiang, Y . Xu, H. Xu, Z. Wang, C. Qiao, and Y . Zhao, “Fedmp: Federated learning through adaptive model pruning in heterogeneous edge computing,” in 2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2022, pp. 767–779

work page 2022
[6]

Group knowledge transfer: Federated learning of large cnns at the edge,

C. He, M. Annavaram, and S. Avestimehr, “Group knowledge transfer: Federated learning of large cnns at the edge,” Advances in Neural Information Processing Systems , vol. 33, pp. 14 068–14 080, 2020

2020
[7]

Federated learning: Challenges, methods, and future directions,

T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated learning: Challenges, methods, and future directions,” IEEE Signal Processing Magazine, vol. 37, no. 3, p. 50–60, May 2020. [Online]. Available: http://dx.doi.org/10.1109/MSP.2020.2975749

arXiv 2020

Show all 32 references

[8]

Federated learning in mobile edge networks: A comprehensive survey,

W. Y . B. Lim, N. C. Luong, D. T. Hoang, Y . Jiao, Y .-C. Liang, Q. Yang, D. Niyato, and C. Miao, “Federated learning in mobile edge networks: A comprehensive survey,” IEEE communications surveys & tutorials , vol. 22, no. 3, pp. 2031–2063, 2020

2020
[9]

Advances and open problems in federated learning,

P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummings et al. , “Advances and open problems in federated learning,” Foundations and trends® in machine learning , vol. 14, no. 1–2, pp. 1–210, 2021

2021
[10]

A survey on federated learning systems: Vision, hype and reality for data privacy and protection,

Q. Li, Z. Wen, Z. Wu, S. Hu, N. Wang, Y . Li, X. Liu, and B. He, “A survey on federated learning systems: Vision, hype and reality for data privacy and protection,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 4, pp. 3347–3366, 2021

2021
[11]

Federated learning with non-iid data,

Y . Zhao, M. Li, L. Lai, N. Suda, D. Civin, and V . Chandra, “Federated learning with non-iid data,” arXiv preprint arXiv:1806.00582 , 2018

2018 arXiv
[12]

Model-contrastive federated learning,

Q. Li, B. He, and D. Song, “Model-contrastive federated learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 10 713–10 722

2021
[13]

Federated optimization in heterogeneous networks,

T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,” Proceedings of Machine learning and systems , vol. 2, pp. 429–450, 2020

2020
[14]

Tackling the objective inconsistency problem in heterogeneous federated optimiza- tion,

J. Wang, Q. Liu, H. Liang, G. Joshi, and H. V . Poor, “Tackling the objective inconsistency problem in heterogeneous federated optimiza- tion,” Advances in neural information processing systems , vol. 33, pp. 7611–7623, 2020

2020
[16]

Dense: Data-free one-shot federated learning,

J. Zhang, C. Chen, B. Li, L. Lyu, S. Wu, S. Ding, C. Shen, and C. Wu, “Dense: Data-free one-shot federated learning,” Advances in Neural Information Processing Systems , vol. 35, pp. 21 414–21 428, 2022

2022
[17]

Data-free knowledge distillation for het- erogeneous federated learning,

Z. Zhu, J. Hong, and J. Zhou, “Data-free knowledge distillation for het- erogeneous federated learning,” in International conference on machine learning. PMLR, 2021, pp. 12 878–12 889

2021
[18]

Towards fair federated learning with zero-shot data augmentation,

W. Hao, M. El-Khamy, J. Lee, J. Zhang, K. J. Liang, C. Chen, and L. C. Duke, “Towards fair federated learning with zero-shot data augmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 3310–3319

2021
[19]

Fine-tuning global model via data-free knowledge distillation for non-iid federated learning,

L. Zhang, L. Shen, L. Ding, D. Tao, and L.-Y . Duan, “Fine-tuning global model via data-free knowledge distillation for non-iid federated learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 10 174–10 183

2022
[20]

Lightsecagg: a lightweight and versatile design for secure aggregation in federated learning,

J. So, C. He, C.-S. Yang, S. Li, Q. Yu, R. E Ali, B. Guler, and S. Avestimehr, “Lightsecagg: a lightweight and versatile design for secure aggregation in federated learning,” Proceedings of Machine Learning and Systems , vol. 4, pp. 694–720, 2022

2022
[21]

Model pruning enables efficient federated learning on edge devices,

Y . Jiang, S. Wang, V . Valls, B. J. Ko, W.-H. Lee, K. K. Leung, and L. Tassiulas, “Model pruning enables efficient federated learning on edge devices,” 2022. [Online]. Available: https://arxiv.org/abs/1909.12326

2022 arXiv
[22]

Dapperfl: Domain adaptive federated learning with model fusion pruning for edge devices,

Y . Jia, X. Zhang, H. Hu, K.-K. R. Choo, L. Qi, X. Xu, A. Beheshti, and W. Dou, “Dapperfl: Domain adaptive federated learning with model fusion pruning for edge devices,” 2024. [Online]. Available: https://arxiv.org/abs/2412.05823

2024 arXiv
[23]

Model compression for communication efficient federated learning,

S. M. Shah and V . K. N. Lau, “Model compression for communication efficient federated learning,” IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 9, pp. 5937–5951, 2023

2023
[24]

Efficient model compression for hierarchical federated learning,

X. Zhu, S. Yu, J. Wang, and Q. Yang, “Efficient model compression for hierarchical federated learning,” 2024. [Online]. Available: https: //arxiv.org/abs/2405.17522

2024 arXiv
[25]

Bayesian federated model compression for communication and computation efficiency,

C. Xia, D. H. K. Tsang, and V . K. N. Lau, “Bayesian federated model compression for communication and computation efficiency,” 2024. [Online]. Available: https://arxiv.org/abs/2404.07532

2024 arXiv
[26]

Learning multiple layers of features from tiny images,

A. Krizhevsky, G. Hinton et al. , “Learning multiple layers of features from tiny images,” 2009

2009
[28]

Measuring the effects of non- identical data distribution for federated visual classification,

T.-M. H. Hsu, H. Qi, and M. Brown, “Measuring the effects of non- identical data distribution for federated visual classification,” arXiv preprint arXiv:1909.06335, 2019

1909 arXiv
[29]

Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,” arXiv preprint arXiv:1708.07747, 2017

2017 arXiv
[30]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

2016
[31]

U-net: Convolutional networks for biomedical image segmentation,

O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 ...

2015
[32]

Rectified linear units improve restricted boltz- mann machines,

V . Nair and G. E. Hinton, “Rectified linear units improve restricted boltz- mann machines,” in Proceedings of the 27th international conference on machine learning (ICML-10) , 2010, pp. 807–814

2010
[33]

Learning to attack federated learning: A model-based reinforcement learning attack framework,

H. Li, X. Sun, and Z. Zheng, “Learning to attack federated learning: A model-based reinforcement learning attack framework,” Advances in Neural Information Processing Systems , vol. 35, pp. 35 007–35 020, 2022

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [27]

Fake It Till Make It: Federated Learning with Consensus-Oriented Generation

R. Ye, Y . Du, Z. Ni, S. Chen, and Y . Wang, “Fake it till make it: Federated learning with consensus-oriented generation,” arXiv preprint arXiv:2312.05966, 2023

work page Pith review arXiv 2023

[2] [1]

Communication-efficient learning of deep networks from decentralized data,

B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Artificial intelligence and statistics . PMLR, 2017, pp. 1273– 1282

2017

[3] [2]

Federated machine learning: Concept and applications,

Q. Yang, Y . Liu, T. Chen, and Y . Tong, “Federated machine learning: Concept and applications,” ACM Transactions on Intelligent Systems and Technology (TIST), vol. 10, no. 2, pp. 1–19, 2019

2019

[4] [3]

Q. Yang, L. Fan, and H. Yu, Federated learning: Privacy and incentive. Springer Nature, 2020, vol. 12500

work page 2020

[5] [4]

Fedpe: Adaptive model pruning-expanding for federated learning on mobile devices,

L. Yi, X. Shi, N. Wang, J. Zhang, G. Wang, and X. Liu, “Fedpe: Adaptive model pruning-expanding for federated learning on mobile devices,” IEEE Transactions on Mobile Computing , 2024

work page 2024

[6] [5]

Fedmp: Federated learning through adaptive model pruning in heterogeneous edge computing,

Z. Jiang, Y . Xu, H. Xu, Z. Wang, C. Qiao, and Y . Zhao, “Fedmp: Federated learning through adaptive model pruning in heterogeneous edge computing,” in 2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2022, pp. 767–779

work page 2022

[7] [6]

Group knowledge transfer: Federated learning of large cnns at the edge,

C. He, M. Annavaram, and S. Avestimehr, “Group knowledge transfer: Federated learning of large cnns at the edge,” Advances in Neural Information Processing Systems , vol. 33, pp. 14 068–14 080, 2020

2020

[8] [7]

Federated learning: Challenges, methods, and future directions,

T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated learning: Challenges, methods, and future directions,” IEEE Signal Processing Magazine, vol. 37, no. 3, p. 50–60, May 2020. [Online]. Available: http://dx.doi.org/10.1109/MSP.2020.2975749

arXiv 2020